(19)
(11) EP 0 466 022 A3

(12) EUROPEAN PATENT APPLICATION

(88) Date of publication A3:
25.08.1993 Bulletin 1993/34

(43) Date of publication A2:
15.01.1992 Bulletin 1992/03

(21) Application number: 91111086.4

(22) Date of filing: 04.07.1991
(51) International Patent Classification (IPC)5: G06F 15/80
(84) Designated Contracting States:
DE FR GB IT

(30) Priority: 12.07.1990 US 551983

(71) Applicant: ALLEN-BRADLEY COMPANY, INC.
Milwaukee Wisconsin 53202 (US)

(72) Inventors:
  • Gasperi, Michael L.
    Racine, Wisconsin 53405 (US)
  • Davis, Wesley
    Franklin, Wisconsin 53132 (US)

(74) Representative: Lippert, Hans, Dipl.-Ing. 
Reichel und Reichel Parkstrasse 13
60322 Frankfurt
60322 Frankfurt (DE)


(56) References cited: : 
   
       


    (54) Teaching method for recurrent neural networks


    (57) A teaching method for a recurrent neural network (10) having hidden (16), output (14) and input (12) neurons calculates weighting errors over a limited number of propagations of the network. This process permits the use of conventional teaching sets, such as are used with feedforward networks, to be used with recurrent networks. The teaching outputs are substituted for the computed activations (Z(3), Z(4)) of the output (14) neurons in the forward propagation and error correction stages. Back propagated error from the last propagation is assumed to be zero for the hidden (16) neurons. A method of reducing drift of the network with respect to a modeled process is also described and a forced cycling method to eliminate the time lag between network input and output.






    Search report